Application of Machine Learning Models for Classifying Wood Surface Defects Using Near-Infrared Spectroscopy

Accurate identification of wood surface defects is crucial for improving the quality and utilization of wood products. To address the low efficiency and accuracy of traditional manual inspection methods, this study collected near-infrared spectroscopy (NIRS) data from Brich and Fir surfaces, including defect-free samples and three typical defect types. The effectiveness of machine learning models in classifying wood surface defects was systematically investigated. Two feature dimensionality reduction methods, principal component analysis (PCA) and recursive feature elimination (RFE), were selected for comparison to screen out representative feature variables. Four classification models, namely, partial least squares discriminant analysis (PLS-DA), random forest (RF), support vector machine (SVM) and fully connected neural network (FCNN), were used to model and classify the wood defect samples. The results indicate that PCA outperforms RFE in enhancing model classification performance. Among the models, the FCNN achieved the best performance, with a highest classification accuracy of 98.85 %, and both recall and F1-score reaching 0.989. These findings demonstrate the superiority of deep learning methods in wood defect recognition tasks. This study systematically evaluated machine learning models based on near-infrared spectroscopy for the classification of wood surface defects, providing valuable insights for model selection and optimization in future research.